Behavior Cloning for Aircraft Autopilots with Semantic Segmentation under Various Lighting Conditions

In this paper, we propose an aircraft autopilot specifically designed for autonomous landing flights. The au- topilot is trained through behavior cloning using a dataset of control command outputs provided by a human pilot in response to image inputs captured from the cockpit view. However, in unknown environments with different lighting conditions, even a trained autopilot struggles to determine appropriate control command outputs for visually different image inputs. To address this challenge and improve general- ization capability across varying lighting conditions, we apply semantic segmentation to the original RGB images to classify runway pixels, and use the resulting segmentation images as inputs to the autopilot. Unlike RGB images, the segmentation images correctly classify only the runway regardless of lighting changes, producing visually consistent representations across all environments. Simulation experiments demonstrate that the proposed autopilot achieves improved generalization compared to the previous RGB-based autopilot, successfully landing on the runway in unknown evening and nighttime environments.